VOLUME 68, NUMBER 19 PHYSICAL REVIEW LETTERS 11 MAY 1992

Using the Sensitive Dependence of Chaos (the "Butterfly Effect") to Direct Trajectories in
an Experimental Chaotic System

Troy Shinbrot, ' ' "' William Ditto Celso Grebogi, ' ' " ' ' Edward Ott
Mark Spano, and James A. Yorke '~ '
"iUniversity of Maryland. College Park, Maryland 20742
Department of Physics, The College of Wooster, Wooster, Ohio 4469l
IVaval Surface Warfare Center. Silver Spring, Maryland 20902
(Received 20 November l99l)
In this paper we present the first experimental verification that the sensitivity of a chaotic system to
small perturbations (the "butterfly effect") can be used to rapidly direct orbits from an arbitrary initial
state to an arbitrary accessible desired state.

PACS numbers: 05.45.+b

Recently, it was demonstrated [I] theoretically and nu-
merically that orbits on a chaotic attractor can be
brought rapidly to a desired state by the application of
tiny, judiciously chosen perturbations to an available sys-
tem parameter. In this Letter we describe the first exper-
imental confirmation of this method.
To illustrate the method in the simplest context (i.e.,
where the attractor dimension is near I), assume that the
dynamics of the system to be controlled are described by
a one-dimensional map [2],
We imagine that the parameter p can be varied by some
small amount about its nominal value P, p P+ tip, and
we seek a value for the small perturbation Bp in some al-
lowed limited range, hp ~8p & hp, which will take us
from a current state, X to a desired state, X, . We ob-
serve that the variation in the state after one iterate of
our map due to the variation in p is
(2)
Since ibpi is restricted to be less than or equal to hp, this
defines an interval ~i. This interval will typically grow
with each successive iteration of the map until it encom-
passes the desired point X, . Once X, is contained within
the interval, we know that some parameter value, pbe-
tween p;p hp and p =p+hp will lead to X&. All
. data points which we fitted with a second spline. Finally,
that remains is to estimate pwhich can be done by a
variety of means.
This procedure will give us a value of p which would, in
the absence of noise or modeling errors, lead us along
some idealized trajectory directly to the target after some
small number of iterations. In a real physical system,
noise and modeling errors will cause the actual trajectory
to wander off the idealized trajectory, however. There-
fore we make periodic corrections by reapplying our tar-
geting algorithm after every iteration. Thus we have a
different value of the parameter on each iterate, n, and
we denote this value p. If the system truly were de-
l992 The American Physical Society
scribed by a one-dimensional map, then our replacement
of p in (I) by pwould be valid. For the system we deal
with, this turns out to be a useful approximation. In gen-
eral, however, the validity of such an approximation has
to be examined on a case-by-case basis. We discuss this
issue at the end of this Letter.
To experimentally evaluate the effectiveness of this
method, we used a vertically oriented, magnetoelastic rib-
bon, which is known [3] to vibrate chaotically in response
to an external applied magnetic field of the form H
,
=Hd, +H, cos(tat). The ribbon was clamped at its base
but was otherwise free to move. The elastic modulus of
the ribbon was nonlinearly dependent on the applied field,
so that as the field oscillated, the ribbon alternately buck-
led and stiffened under the inIIuence of gravity. The posi-
tion X of the ribbon was measured at a point near its base
with an optical sensor.
In order to apply our targeting algorithm, we first con-
structed a map from the experimental system. We chose
[4] to use the dc field Hd, as our control parameter p and
we constructed a map as follows. First, we selected nomi-
nal values of H,. Hdand to, and formed an experimen-
tal delay plot of 500 points by sampling the position of
the ribbon once per driving period, 2tr/ro. This is shown
[5] in Fig. I. We then fitted these points with a robust
spline curve. This gave us a map FP(, X) for a nominal
value p of the parameter. Next, we changed the map by
decreasing the parameter to p hp, obtaining 500 new
we increased the parameter to p+Ap, again obtaining
500 points, and fitted these with a third spline. We used
the spline fits to estimate the value of the map function F
for the three parameter values at any given value of X.
We could then interpolate between these three values to
estimate F(p, X) for any value of the parameter, p, in
our range, [p hp, p+Ap]. Given this model of our
map, we could apply our targeting method in a straight-
forward way to reach a given target point.
To illustrate our procedure, we target the point
X=2.5. We remark that we also targeted other accessi-
ble points. It is interesting to discuss targeting of X =2.5,
2863
VOLUME 68, NUMBER 19 PHYSICAL REVIEW LETTERS 11 MAY 1992

3.5
0 0 0
0 0

C
0 C C C
0
2.5 0
8

1.5 oo
O~ p ~ A 0 C
0 0
Qn 0 Q C 0
0 0
3 0 0 0 0 0 0
0 0 0 0
I I I

0.5 50 100 150

().5

X Iteration, n

FIG. l. Experimental delay plot of magnetoelastic ribbon at

nominal parameter values.

however, because it is in a region of relatively low mea- I l

QP
sure and therefore its vicinity typically takes a long time Lt. &TD
8 ,'j, 'P~(l I
to reach in the absence of our targeting procedure. In CJ

Fig. 2(a), we show a time series of the ribbon position LI II )I

rh
~ IH
I
Il ll
with fixed, nominal, applied field, which clearly shows Q
that the vicinity of X=2.5 is seldom visited. Indeed, for 11 1

each of our three fixed parameter values, p hp, p, and

p+hp, the ribbon position reaches the neighborhood I

10
I

15
l

20 25
[2.49, 2. 51] only about I time in 500 iterates.
By contrast, in Fig. 2(b), we show the results of target- Iteration, n
ing for several representative trajectories. For each tra- FIG. 2. (a) Ribbon displacement vs iteration for fixed, nomi-
jectory shown, we let the ribbon vibrate chaotically for nal parameter values. (b) Typical ribbon displacements vs
100 iterates, and then on the 101st iterate, denoted n= 1, iteration for targeting on. The point X=2.5 is targeted starting
we initiated the targeting algorithm. The position of at iteration n = l.
the ribbon at n=1 is shown as an open circle, and the
position when the ribbon reaches the neighborhood,
[2.49, 2. 51], is shown as a solid circle in the figure. Every tion.
time that targeting is turned on, the ribbon is brought To evaluate the effect of approximating behavior whose
rapidly to the desired target neighborhood [2.49, 2. 51]. dimension is slightly above 1 with a one-dimensional
We reiterate that due to noise and modeling errors, it was map, we implemented our targeting algorithm numerical-
necessary to repeat our targeting algorithm at every itera-
tion. We carried out this procedure for 100 randomly
chosen initial conditions on the attractor, and rapidly
brought the trajectory to the targeted point each and
)

=p+ bX X,
X))
ly on the Henon system defined by the map

every time. On average, we were able to reach the neigh-
borhood [2.49, 2.51] in 20 iterates [6], as compared to 500
iterates that would be required without targeting.
Despite the apparent effectiveness of our targeting al-
gorithm, we encountered two complications which may be
of importance in future applications. First, as we men-
tioned earlier, the technique we used relies on the map
(1) being approximately one dimensional. Yet higher-
dimensional behavior is manifestly present in our system,
as can be seen in Fig. 1 (in particular, note the structure
on the right half of the map indicative of fractal behav-
ior). The targeting algorithm is nevertheless successful.
It is worthwhile, however, to consider the limits of our al-
gorithm due to our use of a one-dimensional approxima-
X+ =F(p, X,
)
for various values of b. We used the parameter p as our
control, and varied p by up to hp =0.05 about the nomi-
nal value p=1.5. We modeled the system as a one-
dimensional map in precisely the same way as in the ex-
perimental system. That is, we generated three data sets
by iterating Eq. (3) using p =1.45, 1.5, and 1.55, and, us-
ing only these data, interpolated between quadratic fits of
these data sets to estimate F(p, X). Using this model,
we chose initial points on the attractor at random [7], and
targeted a particular point, X=1.1, in 25 realizations at
each of several values of the parameter b. The results of
this process are shown in Fig. 3. We show the rms error
of the quadratic fit in the abscissa of the figure [8].

2864
VOLUME 68, NUMBER 19 PH YSICAL REVIEW LETTERS 11 MAY 1992

14 x+i on p~ is also implied by the use of delay coordi-

nates [10].
12- To examine the dependence of x+i on p for
m ~n 1, we performed the following experiment. We
fixed the parameter at the minimum value, hp, p;=p
o 10
for 10 iterates. Then, we switched to the maximum
value, p =p+hp, for 10 more iterates. At this time,
.

0 8- we switched back to We repeated this process for p;.

50 cycles and produced a delay plot of the position of the
ribbon, as shown in Fig. 4(a). For comparison, Fig. 4(b)
shows plots of x~i vs x for the case where p is kept
4
,
fixed at p =p . and for the case where p is kept fixed at
.00 [ p=p~;. For the data in Fig. 4(a), for most of the
RMS Error in 1-0 Fit switches, the subsequent x point does indeed land on the
appropriate approximate curve in Fig. 4(b). However,
FIG. 3. Effect on targeting of rms error due to approximating occasionally, when p is switched from to p p=p;
higher-dimensional attractor by a one-dimensional map.
=p, . , and the p coordinate is in the region la- =p;x
beled 0 in Fig. 4(a), then the subsequent x value occurs

It is clear from Fig. 3 that our procedure can tolerate

in the region labeled 1. Region 1 is not on the p . p,
curve [the lower curve of Fig. 4(b)]. Region 1 in turn
higher dimensionality only until the rms error due to
iterates to region 2 (which also is appreciably far from
higher dimensionality reaches within an order of magni-
tude of the change (+'0.05) which we can produce by
parametrically varying the position on the attractor. In (&) 3.2
our physical experiment, the maximum change [9] in the 3.0- ~ +g \

position on the attractor which we produced by varying +e' ~q+g

'4 W
2. 8- ~ 4

the parameter amounted on average to 5% of the position

itself. By constrast, the rms error in our spline fits was 2.6-

about 1% of the position. According to our simulation of t.~

the effect of higher dimensionality, our experiment is ap- C
2.2
parently operating very near the margin of targeting
effectiveness. Indeed, our targeting algorithm was ob- 2.0-

served in the experiment to be substantially less effective i.s- 2

for smaller parameter variations. 1.6
The second complication has to do with the fact that in
our analysis we have assumed that the system dependence 1.6
I I

1.8
I

2.0
I

2.2
I

2. 4
I

2.6
I

2.H
I

3.0 3.2
on the parameter is given by x+~ =F(x,p) with F a
xn
one-dimensional map. We note, however, that in this pa-
per we are dealing with a surface of section for a continu- (b) 32
ous time (t), infinite-dimensional system. Under such 3.0-
~ P
circumstances, even if a one-dimensional map applies at
each fixed parameter value, and even if contraction to the
" a~~
a, 's,
~=~mn
26
attractor is very rapid compared to the surface of section
map iteration time, ht, then the switching of p[ to p in f.
principle leads to a dependence of x+ on both p i and ~
22 '~l

p. To see this, we note that the attractor position in the

full infinite-dimensional phase space depends on p. A
1.8-
point with scalar coordinate xat
time n is on the
p =p attractor at time t
~
=t.
At a time shortly after 1.6

the switch from p[ to p it is essentially on the p =p 1.4 I I I I I I I

attractor (assuming rapid contraction), and in general 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2

now has moved to a new location that has an x coordi- Xn

nate, x=g(x,p,p-i), that depends on the location of FIG. 4. (a) Delay plot for a parameter alternately switched
the attractors for p =p ] and for p =p. The location of between p~;, and p,
. The nominal parameters HdH, and
x+~ is determined by iterating the point xvia the map co used in this experiment are slightly different from those in
that applies for p fixed for all time at p. Thus x~] de- Fig. l. (b) Delay plots for a parameter fixed at and for a p;,
pends on both pand pi. An added dependence of parameter fixed at . p,
2865
VOLUME 68, NUMBER 19 PHYSICAL REVIEW LETTERS
the appropriate curve) and thereafter rapidly approaches
the p =p .curve. While this effect degrades the target-
ing performance, it occurs relatively rarely. Presumably,
a more elaborate procedure could, in principle, correct for
this, but we have not found such a procedure to be neces-
sary.
In conclusion, this experiment has confirmed for the
first time that it is possible to direct a chaotic system
from an arbitrary initial state to a particular desired state
using only small perturbations to an accessible system pa-
rameter. The method used is particularly suitable for
systems which are describable by a one-dimensional map,
although some higher dimensionality can be tolerated.
We stress that we have achieved control of an experimen-
tal system in the absence of an a priori theoretical model
[I I].
The authors wish to acknowledge the assistance and
helpful comments of Dr. M. Wun-Fogle of the NSWC.
This research was supported by the U. S. Department of
Energy (Scientific Computing Staff Office of Energy
Research).
"'
Institute for Physical Science and Technology.
Department of Physics.
' Laboratory for Plasma Research.
Department of Mathematics.
' Department of Electrical Engineering.
[I] T. Shinbrot, E. Ott, C. Grebogi, and J. A. Yorke, Phys.
Rev. Lett. 65, 3215 (1990).
[2] T. Shinbrot, E. Ott, C. Grebogi, and J. A. Yorke, Phys.
Rev. A 45, 4165 (1992). This reference considers chaotic
systems which are describable by a one-dimensional map.
2866
11 MAY 1992
Reference [I] considers chaotic systems ~hose fractal di-
mension is between 1 and 2 and which cannot be approxi-
mately described by a one-dimensional map.
[3] W. L. Ditto, S. Rauseo, R. Cawley, C. Grebogi, G. -H.
Hsu, E. Kostelich, E. Ott, H. T. Savage, R. Segnan, M.
L. Spano, and j. A. Yorke, Phys. Rev. Lett. 63, 923
(1989).
[4] The other parameters Hor ai were tried as well. Our
method worked about equally well with either ac field H, ,
or dc field Hd, . We wanted to avoid changing the fre-
quency since we sampled the ribbon position once per fre-
quency cycle.
[5] It is worth mentioning that we sampled the position of the
ribbon at zero phase of the sinusoidal driving function.
Since the experimental system is continuous, we could in-
stead have sampled the position at any other phase. Our
delay plot would then have looked different, but our tech-
nique would work equally well.
[6] In our 100 realizations, we observed a few outlying cases
(-50 iterates).
requiring abnormally long targeting times
Deletion of these few cases from the record results in a
significantly shorter average targeting time (-10 iter-
ates).
[7] To choose random initial points, we iterated Eq. (3) 25
times using p=1.5, starting from X, -] =0 and using X,
randomly distributed on [ 0.5, 0.5].
[8] To aid the eye, we have also shown in the figure a best-fit
hyperbolic tangent obtained by nonlinear regression.
[9l This is the average of the difference between robust spline
positions from the three curves divided by the nominal po-
sition. We normalized the position of the ribbon to a
scale from 0 to 1.
[10] U. Dressier and G. Nitsche (to be published).
[I I] Experimental targeting for attractors of higher dimen-
sionality (as in our numerical work of Ref. [I]) remains a
problem for future study.